Hu Zixuan, Lin Hui Ming, Mathur Shobhit, Moreland Robert, Witiw Christopher D, Jimenez-Juan Laura, Callejas Matias F, Deva Djeven P, Sejdić Ervin, Colak Errol
The Edward S. Rogers Department of Electrical and Computer Engineering, University of Toronto, Toronto, ON, Canada.
Department of Medical Imaging, St Michael's Hospital, Unity Health Toronto, 30 Bond St, Toronto, ON, M5B 1W8, Canada.
NPJ Digit Med. 2025 May 7;8(1):254. doi: 10.1038/s41746-025-01594-2.
This study proposes a semi-weakly supervised learning approach for pulmonary embolism (PE) detection on CT pulmonary angiography (CTPA) to alleviate the resource-intensive burden of exhaustive medical image annotation. Attention-based CNN-RNN models were trained on the RSNA pulmonary embolism CT dataset and externally validated on a pooled dataset (Aida and FUMPE). Three configurations included weak (examination-level labels only), strong (all examination and slice-level labels), and semi-weak (examination-level labels plus a limited subset of slice-level labels). The proportion of slice-level labels varying from 0 to 100%. Notably, semi-weakly supervised models using approximately one-quarter of the total slice-level labels achieved an AUC of 0.928, closely matching the strongly supervised model's AUC of 0.932. External validation yielded AUCs of 0.999 for the semi-weak and 1.000 for the strong model. By reducing labeling requirements without sacrificing diagnostic accuracy, this method streamlines model development, accelerates the integration of models into clinical practice, and enhances patient care.
本研究提出了一种用于在CT肺动脉造影（CTPA）上检测肺栓塞（PE）的半弱监督学习方法，以减轻详尽医学图像标注所需的资源密集型负担。基于注意力的CNN-RNN模型在RSNA肺栓塞CT数据集上进行训练，并在一个汇总数据集（Aida和FUMPE）上进行外部验证。三种配置包括弱监督（仅检查级标签）、强监督（所有检查和切片级标签）和半弱监督（检查级标签加有限的切片级标签子集）。切片级标签的比例从0到100%不等。值得注意的是，使用大约四分之一的总切片级标签的半弱监督模型的AUC为0.928，与强监督模型的AUC 0.932非常接近。外部验证得出半弱监督模型的AUC为0.999，强监督模型的AUC为1.000。通过在不牺牲诊断准确性的情况下减少标注要求，该方法简化了模型开发，加速了模型在临床实践中的整合，并改善了患者护理。
